What I'm Building

Real-Estate SaaS Rebuild — 3 Days, Spec to Working Application

The problem: I had a working prototype with real-user feedback — and that feedback revealed the architecture needed to be replaced, not patched.

What I did: Authored a detailed specification covering the existing functionality and the proposed changes, handed it to Claude Code, and got a one-shot rebuild that recovered all prior functionality and implemented most of the new features. A few focused sessions over the following days finished the rest.

Total elapsed time: three days. That same scope would have taken a solo developer roughly six months two years ago.

What this demonstrates: The concrete productivity delta when disciplined spec-writing meets agent orchestration. The skill isn't typing faster — it's knowing exactly what to build and how to direct the work to get there.

AI Content Scoring Platform

The problem: Processing content at scale needs different models for different tasks — and the costs add up fast if you're not deliberate about which model handles what.

What I built: A content scoring platform with multi-LLM provider architecture across OpenAI, Anthropic, Google Gemini, and local quantized models — cost-optimized routing, automatic provider fallback, and a production-grade API for bulk processing.

The differentiator: I ran a 6-model bake-off measuring speed, cost, and quality on real workloads. Gemini 2.5 Lite won the cost-quality envelope for text analysis — a contrarian result that beat the "always reach for the frontier model" default.

What this demonstrates: I choose models empirically, by measuring them on the actual work — not by chasing frontier-model headlines.

Multi-Agent Document Processing Pipeline

What I built: A production pipeline that classifies inputs, routes them to specialized generation agents based on extracted criteria, and enforces quality constraints at agent boundaries — verbatim source preservation, no-fabrication rules, traceable provenance — producing tier-routed output at scale.

Persistent state, batch processing of 100+ items, human-in-the-loop checkpoints, and a cross-run memory layer for pattern learning. The use case is application-specific; the architecture is general.

What this demonstrates: Agent orchestration as an engineering discipline — quality enforcement and state management, not just prompting.

"Open Mic" — Adaptive Lyric Scroller (iOS)

The problem: Musicians at open mic nights can't scroll lyrics at a fixed speed. Tempo shifts, dramatic pauses, and personal habits make every performance different. Existing teleprompter apps either scroll too fast or too slow — and you can't exactly reach for your phone mid-song.

What I'm building: A SwiftUI app that listens to live performance audio, determines the playing tempo in real-time, and adjusts lyric/chord scrolling speed accordingly. Users "train" the app across practice sessions so it learns their individual timing on specific songs.

This combines real-time audio analysis, on-device machine learning, and adaptive UX design — and it started because I go to a lot of open mic nights and watched musicians struggle with this exact problem.

Status: In active development.

What this demonstrates: Product thinking. I identify real problems that real people have, then figure out whether technology can solve them. Sometimes it can. Sometimes the answer is "just memorize the lyrics" — knowing the difference is the skill.

Custom RAG Pipeline via MCP

A tool I reach for in specific situations — not a headline capability.

The problem: I needed to reference thousands of pages of developer documentation while coding, but it only existed as web pages, and switching between browser tabs and my editor was killing my workflow.

What I built: A pipeline that ingests thousands of documentation pages, processes them for retrieval, and serves them through a Model Context Protocol (MCP) server — so I can query documentation in real-time from inside my development environment.

The honest framing: RAG and MCP were the headline of early-2024 AI. Today I treat them as specific-situation tools, applied project-by-project after measuring cost/benefit. I built this because I had a real problem — not because "RAG" looks good on a resume.